Simulation of NMR response from micro-CT images using artificial neural networks

The Nuclear Magnetic Resonance (NMR) log is amongst the functional techniques in petroleum investigation to segregating the reservoir and non-reservoir horizons precisely; furthermore, the NMR log provides an improved method to determine reservoir petrophysical parameters. Unfortunately, these data are usually sparse since acquiring NMR logs in producing cased wells is not possible and it is one of the most expensive tools in the logging industry thus its associated costs are the major limitation of its usage. Consequently, researchers have recently studied to virtually extract the NMR parameters via other routes. One such route, which we propose here is the possibility of estimating the T2 distribution curve and magnetization decay by establishing a relationship between micro-CT images and NMR parameters by means of artificial neural networks (ANN) and image analysis algorithms. Specifically, two ANN networks were designed, which numerically image features from micro-CT images as inputs, while the amplitude of the magnetization and relaxation time were output parameters. We assessed the procedure by taking the error rate and correlation coefficient into consideration and we conclude that the ANN model is capable of finding logical patterns between image features and NMR responses, and is thus able to reliably predict NMR response behavior. Furthermore, we quantitatively compared ANN and random walk (RW) NMR predictions, and we demonstrate that ANN readily outperforms RW in terms of accuracy.